CN108932499A - A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network - Google Patents

A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network Download PDF

Info

Publication number
CN108932499A
CN108932499A CN201810742624.9A CN201810742624A CN108932499A CN 108932499 A CN108932499 A CN 108932499A CN 201810742624 A CN201810742624 A CN 201810742624A CN 108932499 A CN108932499 A CN 108932499A
Authority
CN
China
Prior art keywords
rolling bearing
network
dbn
rbm
dimensional
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201810742624.9A
Other languages
Chinese (zh)
Inventor
蒋黎明
徐春玲
李友荣
徐增丙
周明乐
聂婉琴
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University of Science and Engineering WUSE
Wuhan University of Science and Technology WHUST
Original Assignee
Wuhan University of Science and Engineering WUSE
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University of Science and Engineering WUSE filed Critical Wuhan University of Science and Engineering WUSE
Priority to CN201810742624.9A priority Critical patent/CN108932499A/en
Publication of CN108932499A publication Critical patent/CN108932499A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G06F2218/10—Feature extraction by analysing the shape of a waveform, e.g. extracting parameters relating to peaks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/467—Encoded features or binary features, e.g. local binary patterns [LBP]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/06—Recognition of objects for industrial automation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Image Analysis (AREA)

Abstract

The present invention proposes a kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network, one-dimensional bearing vibration signal is extracted first, it is then converted into two-dimensional time-domain gray scale picture, and to two-dimensional time-domain gray scale picture piecemeal, block diagram is utilized respectivelyExtract Local textural feature, after each block diagram is obtainedStatistic histogram be combined and the input as DBN, DBN automatically extracts the further feature of statistic histogram, and by, to self study and the backpropagation declined based on gradient, adjusting the model parameter of entire DBN network before DBN, the DBN network of training completion is obtained；Finally, the vibration signal two dimensional gray figure of the rolling bearing of unknown state is passed throughStatistic histogram feature as input, the high-level characteristic for being able to reflect extrinsic information is successively extracted using the DBN network that training is completed, layer-by-layer feature is input to top-level categories from the result extracted again, realizes fault identification of rolling bearing in the case where multi-load and very noisy.

Description

A kind of rolling bearing prediction technique based on local binary patterns and depth confidence network And system
Technical field
The invention belongs to image procossings, mechanical equipment health detection field, are related to local binary patterns algorithm, in particular to A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network.
Background technique
Rolling bearing is most common in rotating machinery and is very crucial mechanical part, is widely used in family and industry In equipment.Since bearing usually works in severe working environment, they are easy to break down at work.If Failure is found not in time, may result in machine hang-up, even results in catastrophic damage.Therefore, it is necessary to take one The health status of kind of detection mode detection rolling bearing, identify whether to break down and the type of failure and failure it is serious Degree, and then necessary measure is taken, prevent the further damage of bearing, it is ensured that equipment is timely rested and reorganized, safe operation.
In recent years with the development of machine learning and other technologies, more and more developers by neural network, support The machine learning techniques such as vector machine, K- nearest neighbor algorithm apply in rolling bearing fault diagnosis.But these technologies all need Manually to extract the feature of bearing vibration signal, this process not only needs technical staff to possess relevant signal profession to know That knows subjectivity goes selection feature, and consumes the plenty of time.Depth confidence network (DBN) is one that Hinton was introduced in 2006 Kind deep neural network, it forms more abstract high level by combination low-level image feature and indicates, it has been found that data distribution formula feature It indicates, is that one kind can be directly from bottom original signal, successively greedy study obtains high-level characteristic learning network.With tradition Method is compared, and depth confidence network combines the more abstract high level of low-level image feature formation by building multitiered network automatically and indicates, Automatically feature is extracted from bottom to high level, and then promotes the accuracy of classification and prediction.
It can be presented in the form of two-dimensional time-domain picture format in certain situation bearing vibration signals.Two-dimensional time-domain figure The relationship that can reflect signal amplitude and time includes more information than one-dimensional signal.In essence, rolling bearing Fault diagnosis is a kind of Fault Pattern Recognition process, closely similar with image recognition processes.Failure modes and image recognition all belong to In the scope of pattern-recognition.Therefore, the calculation method of field of image processing is introduced into the fault diagnosis of rolling bearing has Higher feasibility.However bearing vibration time-domain image is directly directly done to the input of DBN, state recognition performance can under Drop.Because the dimension of X-Y scheme be it is huge, directly as the input of DBN, will cause dimension disaster, consume a large amount of time. And DBN can ignore the local feature of image, and the local feature of the two-dimensional time-domain figure of rolling bearing often includes abundant Fault state information.Therefore it needs to automatically extract the method for image procossing and DBN the efficient recognition capability of feature to tie It closes.Current most image method such as Scale invariant features transform (SIFT) needs artificial setup parameter, and to two dimension Image dimensionality reduction effect needs to improve.
In view of the defect of appeal, the present invention proposes a kind of axis of rolling based on local binary patterns and depth confidence network Prediction technique is held, the deficiency of conventional fault diagnosis method is solved.
Summary of the invention
The technical issues of technical solution provided according to embodiments of the present invention solves is traditional bearing diagnostic method and image Processing method combination effect is bad, it is difficult to carry out preferable feature extraction and dimensionality reduction, mining data sample to vibration time domain X-Y scheme This essential information, nicety of grading is limited and calculates the larger problem of consumption.
The present invention relates to unified local binary patterns and depth confidence network, basic principle are as follows:
1.LBP (i.e. Local Binary Pattern) is initially as a kind of auxiliary square for handling topography's contrast Method.It has the advantages that carry out dimensionality reduction to dimensional images and rotational invariance and gray scale invariance etc. are significant, the zone of action For a fritter local feature of textural characteristics, and mark the feature of Pixel-level.
LBP is defined on 3 × 3 any radius and any neighborhood neighborhood of a point, mark LBPP,R, P is neighborhood points, and R is The radius of neighbourhood.The basic handling method of LBP is, to some (not including edge pixel point) pixel of original image, pcAnd pcAround 8 pixel (pi, i=1,2 ..., 8) it is formed.With central pixel point PcGray value be threshold value, by pcAdjacent the 8 of surrounding The gray value and threshold value p of a pixelcIt is successively compared, as neighborhood territory pixel gray value pi≥pcWhen, corresponding position is encoded to 1, it is otherwise 0.Then 8 bits are read (assuming that from p according to clockwise direction0Start) to get arrive central pixel point pc LBP value (range is 0-255).Above description is formulated are as follows:
Wherein, P and R indicates the quantity of neighbor pixel and the radius of neighborhood, pcIndicate center pixel (xc,yc) gray scale Value, piAnd indicate the gray level of ith pixel in neighborhood.
The study found that only a small number of partial modes describe the important model of image texture, the probability occurred reaches 90% More than, this mode is exactly " More General Form ".If binary system is regarded as a circle, More General Form be exactly in string from 0 to 1 and from 1 to 0 conversion is no more than mode twice, mark(0 transformation), 01110000 (2 times for example, mode 00000000 Transformation), 11001111 (2 times transformation) be More General Form, referred to as consistent LBP.And 11001001 (4 transformations), 01010011 (6 transformations) is not More General Form, referred to as non-uniform LBP.With P=8, for R=1, LBPP,RMode have 256 kinds, andMode at most only have 59 kinds, as shown in figure 3, it is anti-although More General Form LBP is only the fraction of LBP output Most texture informations have been reflected, still there is very strong descriptive power.Above description is formulated are as follows:
Wherein, P and R indicates the quantity of neighbor pixel and the radius of neighborhood, pcIndicate the gray value of center pixel, piAnd And indicate the gray level of ith pixel in neighborhood.
2. depth confidence network
Depth confidence network is substantially the back propagation network for having supervision by multiple limited Boltzmann machines (RBM) and one layer Multi-layered perception neural networks made of network (BPNN) stacks, each limited Boltzmann machine is made of visual layers and hidden layer, Depth confidence network as shown in Figure 3 includes 3 layers of RBM.
The first layer: the first visual layers v1For original input data and the first hidden layer h1Form first RBM (RBM1)；
The second layer: the first hidden layer h1As the second visual layers v2, and with the second hidden layer h2Form second RBM (RBM2)；
The third time: the second hidden layer h2As third visual layers v3, and with third layer hidden layer h3Form third RBM(RBM3)
It is connected between layers by weight w.It is mutually indepedent inside each layer, it is each hidden when the state of given visible element Hiding between layer state of activation is conditional sampling, at this point, the activation probability of j-th of hidden unit are as follows:
WhereinFor sigmoid activation primitive, θ={ wij,ai,bjBe RBM parameter, wijIndicate visible Connection weight between unit i and hidden unit j, aiIndicate the biasing of visual element i, bjIndicate the biasing of Hidden unit j, hj It is j-th of hidden unit state, viIt is i-th of visible element state, there are two types of their states: 0 or 1.Due to the knot of RBM Structure is symmetrically, when the state of given hidden unit, it is seen that the activation condition of unit is also conditional sampling, i.e., i-th can Depending on the activation probability of unit are as follows:
RBM is the stochastic neural net that an activation primitive is sigmoid function, is trained by way of iteration To parameter θ={ wij,ai,bjAs a result, and being fitted with given training data.Optimized parameter θ*It can be by training set Very big log-likelihood function obtain, i.e.,
Wherein, t is the number of iterations, and T is maximum number of iterations.
In order to quickly calculate the log-likelihood gradient of RBM, weight and biasing can be obtained using the algorithm to sdpecific dispersion Parameter updates, and indicates are as follows:
Δωij=ε (< vihj>data-<vihj>recon)
Δai=ε (< vi>data-<vi>recon)
Δbi=ε (< hj>data-<hj>recon)
Wherein: ε is the learning rate of pre-training,<>dataFor mathematic expectaion defined in training data,<>reconAttach most importance to Mathematic expectaion defined in the model of structure.
BP network is a kind of classifier of supervised learning, can be classified to the feature vector that RBM pre-training obtains, and And play the role of the entire DBN parameter of fine tuning.The training of BP network is divided into two stages of propagated forward and back-propagating.
The propagated forward stage: input feature value is successively traveled to output layer, the classification results predicted.
The reversed fine tuning stage: this process is gradually to finely tune model to bottom using known label from DBN the last layer Parameter, referred to as after to fine tuning learn.For network the last layer by Soft-max as classifier, Softmax model is substantially more The Logic Regression Models of classification.If DBN is stacked by each RBM of l altogether, initial sample x, the last layer output vector is ul (x)
Wherein, blFor the biasing of l layers of RBM, wlFor the weight of l layers of RBM, i-th of sample is folded under RBM through preceding to l layer heap After study, belong to classification yi, yiThe probability of ∈ (1,2 ..., c) is
V is parameter coefficient, and choosing classification corresponding to maximum probability is that Soft-max model determines classification.
L layers of error function express formulas
ρ is weight attenuation rate, λl={ wl,bl,cl,Vl, 1 { yi=k } it is logic display function, work as yiWhen=k, being worth is 1. Work as yiWhen ≠ k, be worth be for 0, m hidden layer neuron number, c be classification number.To seek error minimum value, risen using gradient Method seeks local derviation to parameter
Finely tune parameter
A is learning rate, and so on, l layers of relevant parameter are arrived in fine tuning 1.
Conceived according to foregoing invention, the present invention adopts the following technical solutions: one kind is set based on local binary patterns and depth Local binary patterns and depth confidence network are merged, not only solve image by the rolling bearing prediction technique of communication network Processing and the bad problem of DBN combination effect are more faced directly and solve the problems, such as image high dimensional data dimensionality reduction and the vibration of rolling bearing two dimension Dynamic temporal signatures extract difficult problem, the specific steps are as follows:
Step 1 pre-processes the vibration signal of the one-dimensional rolling bearing of known state, obtains the rolling of known state The two-dimensional time-domain grayscale image of dynamic bearing；
The two-dimensional time-domain gray level image of the rolling bearing of obtained known state is carried out region division, divided by step 2 Principle are as follows: take the common divisor of picture traverse and height respectively；
Above-mentioned gray level image is divided region and used based on unified local binary patterns by step 3Operator carries out special Sign is extracted, and is obtained each statistic histogram for dividing region, then the statistic histogram that each division region obtains is combined, is obtained The statistic histogram of whole picture figure, and statistic histogram is normalized；
The statistic histogram of entire image is input to DBN network (depth confidence network) by step 4, and to DBN network It is handled before carrying out to self study and based on the backpropagation processing that gradient declines, adjusts the model parameter of DBN network, trained The DBN network of completion；
Step 5, when identifying practical multi-load rolling bearing fault mode, first to the rolling bearing of unknown state Vibration signal carries out the processing of step 1 to three, obtains its corresponding statistic histogram, is then completed using obtained training DBN network classify, so that it is determined that the fault type of practical rolling bearing.
Further, in the step 1, the vibration signal of the one-dimensional rolling bearing to known state is normalized Processing makes its scope control in [0,1], the vibration signal after normalization is converted to two dimensional gray time-domain diagram, picture later Amplitude be set as [0-1]；It is required that being converted into the vibration signal length of the one-dimensional rolling bearing of two dimensional gray time-domain diagram should meet It turns around the number of the acquired data point of sensor greater than shaft；The resolution ratio of two dimensional gray figure scale having the same.
Further, in the step 2, region is carried out to the two-dimensional time-domain gray level image of the rolling bearing of known state The Specific Principles of division are as follows: divided according to the resolution sizes of two dimensional gray figure, if the width of resolution ratio and height are n Multiple, then two dimensional gray figure is blocked into n × n；And the more meeting discriminations of piecemeal are higher, but operation time will will increase.
Further, DBN network described in step 4 is substantially and is had by multiple limited Boltzmann machines (RBM) and one layer Multi-layered perception neural networks made of the counterpropagation network (BPNN) of supervision stacks, specific structure is as follows,
Each limited Boltzmann machine is made of visual layers and hidden layer, is connected between layers by weight w, each layer It is internal mutually indepedent, it is conditional sampling between each hidden layer state of activation, at this point, jth when the state of given visible element The activation probability of a hidden unit are as follows:
WhereinFor sigmoid activation primitive, θ={ wij,ai,bjBe RBM parameter, wijIndicate visible Connection weight between unit i and hidden unit j, aiIndicate the biasing of visual element i, bjIndicate the biasing of Hidden unit j, hj It is j-th of hidden unit state, viIt is i-th of visible element state, there are two types of their states: 0 or 1；Due to the knot of RBM Structure is symmetrically, when the state of given hidden unit, it is seen that the activation condition of unit is also conditional sampling, i.e., i-th can Depending on the activation probability of unit are as follows:
RBM is the stochastic neural net that an activation primitive is sigmoid function, is trained by way of iteration To parameter θ={ wij,ai,bjAs a result, and being fitted with given training data, optimized parameter θ*It can be by training set Very big log-likelihood function be
In order to quickly calculate the log-likelihood gradient of RBM, weight and biasing can be obtained using the algorithm to sdpecific dispersion Parameter updates, and indicates are as follows:
Δωij=ε (< vihj>data-<vihj>recon)
Δai=ε (< vi>data-<vi>recon)
Δbi=ε (< hj>data-<hj>recon)
Wherein: ε is the learning rate of pre-training,<>dataFor mathematic expectaion defined in training data,<>reconAttach most importance to Mathematic expectaion defined in the model of structure；
BP network is a kind of classifier of supervised learning, can be classified to the feature vector that RBM pre-training obtains, and And play the role of the entire DBN parameter of fine tuning, the training of BP network is divided into two stages of propagated forward and back-propagating；
The propagated forward stage: input feature value is successively traveled to output layer, the classification results predicted；
The reversed fine tuning stage: the classification results that prediction obtains are compared to obtain error with standard markup information, will accidentally Difference successively returns backward, to realize the fine tuning of DBN parameter；
If DBN is stacked by l RBM altogether, initial sample x, the last layer output vector is ul(x),
Wherein, blFor the biasing of l layers of RBM, wlFor the weight of l layers of RBM, i-th of sample is folded under RBM through preceding to l layer heap After study, belong to classification yi, yiThe probability of ∈ (1,2 ..., c) is
V is parameter coefficient, and choosing classification corresponding to maximum probability is that Soft-max model determines classification；
L layers of error function express formulas
ρ is weight attenuation rate, λl={ wl,bl,cl,Vl, 1 { yi=k } it is logic display function, work as yiWhen=k, being worth is 1, Work as yiWhen ≠ k, be worth be for 0, m hidden layer neuron number, c be classification number；To seek error minimum value, risen using gradient Method seeks local derviation to parameter
Finely tune parameter
A is learning rate, and so on, l layers of relevant parameter are arrived in fine tuning 1.
In addition, the present invention provides a kind of rolling bearing forecasting system based on local binary patterns and depth confidence network, Including following module:
Preprocessing module, the vibration signal for the one-dimensional rolling bearing to known state pre-process, and obtain known The two-dimensional time-domain grayscale image of the rolling bearing of state；
Image division module, for the two-dimensional time-domain gray level image of the rolling bearing of obtained known state to be carried out region It divides, division principle are as follows: take the common divisor of picture traverse and height respectively；
Characteristic extracting module is used for above-mentioned gray level image to be divided region based on unified local binary patternsOperator carries out feature extraction, obtains each statistic histogram for dividing region, then the statistics histogram that each division region is obtained Figure is combined, and obtains the statistic histogram of whole picture figure, and statistic histogram is normalized；
DBN network training module, for the statistic histogram of entire image to be input to DBN network (depth confidence net Network), and the backpropagation for handling before carrying out to DBN network to self study and being declined based on gradient is handled, and the mould of DBN network is adjusted Shape parameter obtains the DBN network of training completion；
Failure modes module is used for when identifying practical multi-load rolling bearing fault mode, first to unknown state The vibration signal of rolling bearing carries out the processing of step 1 to three, obtains its corresponding statistic histogram, then using acquired Training complete DBN network classify, so that it is determined that the fault type of practical rolling bearing.
Further, in the preprocessing module, the vibration signal of the one-dimensional rolling bearing to known state is returned One change processing, makes its scope control in [0,1], the vibration signal after normalization is converted to two dimensional gray time-domain diagram later, The amplitude of picture is set as [0-1]；It is required that the vibration signal length for being converted into the one-dimensional rolling bearing of two dimensional gray time-domain diagram is answered Meet and is greater than shaft and turns around the number of the acquired data point of sensor；The resolution ratio of two dimensional gray figure scale having the same.
Further, in described image division module, to the two-dimensional time-domain gray level image of the rolling bearing of known state into The Specific Principles of row region division are as follows: divided according to the resolution sizes of two dimensional gray figure, if the width of resolution ratio and height Degree is the multiple of n, then two dimensional gray figure is blocked into n × n；And the more meeting discriminations of piecemeal are higher, but operation time will Increase.
Further, DBN network described in DBN network training module is substantially by multiple limited Boltzmann machines (RBM) and one layer has multi-layered perception neural networks made of counterpropagation network (BPNN) stacking of supervision, and specific structure is as follows,
Each limited Boltzmann machine is made of visual layers and hidden layer, is connected between layers by weight w, each layer It is internal mutually indepedent, it is conditional sampling between each hidden layer state of activation, at this point, jth when the state of given visible element The activation probability of a hidden unit are as follows:
WhereinFor sigmoid activation primitive, θ={ wij,ai,bjBe RBM parameter, wijIndicate visible Connection weight between unit i and hidden unit j, aiIndicate the biasing of visual element i, bjIndicate the biasing of Hidden unit j, hj It is j-th of hidden unit state, viIt is i-th of visible element state, there are two types of their states: 0 or 1；Due to the knot of RBM Structure is symmetrically, when the state of given hidden unit, it is seen that the activation condition of unit is also conditional sampling, i.e., i-th can Depending on the activation probability of unit are as follows:
RBM is the stochastic neural net that an activation primitive is sigmoid function, is trained by way of iteration To parameter θ={ wij,ai,bjAs a result, and being fitted with given training data, optimized parameter θ*It can be by training set Very big log-likelihood function be
In order to quickly calculate the log-likelihood gradient of RBM, weight and biasing can be obtained using the algorithm to sdpecific dispersion Parameter updates, and indicates are as follows:
Δωij=ε (< vihj>data-<vihj>recon)
Δai=ε (< vi>data-<vi>recon)
Δbi=ε (< hj>data-<hj>recon)
Wherein: ε is the learning rate of pre-training,<>dataFor mathematic expectaion defined in training data,<>reconAttach most importance to Mathematic expectaion defined in the model of structure；
BP network is a kind of classifier of supervised learning, can be classified to the feature vector that RBM pre-training obtains, and And play the role of the entire DBN parameter of fine tuning, the training of BP network is divided into two stages of propagated forward and back-propagating；
The propagated forward stage: input feature value is successively traveled to output layer, the classification results predicted；
The reversed fine tuning stage: the classification results that prediction obtains are compared to obtain error with standard markup information, will accidentally Difference successively returns backward, to realize the fine tuning of DBN parameter；
If DBN is stacked by l RBM altogether, initial sample x, the last layer output vector is ul(x),
Wherein, blFor the biasing of l layers of RBM, wlFor the weight of l layers of RBM, i-th of sample is folded under RBM through preceding to l layer heap After study, belong to classification yi, yiThe probability of ∈ (1,2 ..., c) is
V is parameter coefficient, and choosing classification corresponding to maximum probability is that Soft-max model determines classification；
L layers of error function express formulas
ρ is weight attenuation rate, λl={ wl,bl,cl,Vl, 1 { yi=k } it is logic display function, work as yiWhen=k, being worth is 1, Work as yiWhen ≠ k, be worth be for 0, m hidden layer neuron number, c be classification number；To seek error minimum value, risen using gradient Method seeks local derviation to parameter
Finely tune parameter
A is learning rate, and so on, l layers of relevant parameter are arrived in fine tuning 1.
The present invention has the effect that compared with existing conventional rolling bearing fault diagnosis technology
The present invention combines the depth confidence network in the local binary patterns and deep learning of image processing techniques simultaneously It applies in rolling bearing fault diagnosis, more enough two-dimension vibration time-domain diagrams to rolling bearing carry out dimensionality reduction and from local grain Useful feature is extracted, and carries out further feature extraction using DBN self-learning capability and obtains higher fault recognition rate, Strengthen fault diagnosis to multi-load case robustness, and reduces computing resource consumption.
Detailed description of the invention
Fig. 1 is the method for the present invention flow chart；
Fig. 2 is LBP algorithm flow chart in the embodiment of the present invention；
Fig. 3 is the two-dimension vibration time domain grayscale image piecemeal of known state and to obtain statistic histogram in the embodiment of the present invention and show It is intended to；
Fig. 4 is the new statistic histogram in the embodiment of the present invention after the combination of block statistics histogram；
Fig. 5 is DBN network model framework figure in the embodiment of the present invention；
Fig. 6 is testing stand schematic diagram provided in an embodiment of the present invention.
Specific embodiment
Below in conjunction with attached drawing to a preferred embodiment of the present invention will be described in detail, it should be understood that described below is excellent It selects embodiment to be only used for describing and explaining invention, is not intended to limit the present invention.
Fig. 1 is the rolling bearing fault mould provided in an embodiment of the present invention based on local binary patterns and depth confidence network Formula recognition methods block diagram, as shown in Figure 1, step includes:
Step 1 handles the vibration signal of the one-dimensional rolling bearing of known state, obtains the rolling of known state The two-dimensional time-domain grayscale image of bearing.The vibration signal of one-dimensional rolling bearing is normalized first, its amplitude is limited in [0,1], Vibration signal is divided into multiple samples according to certain length later, sample is changed into the gray scale of identical size respectively, is schemed In amplitude be limited to [0-1].
The two-dimensional time-domain gray level image of the rolling bearing of the obtained known state is carried out region division by step 2 (as shown in figure 3, being divided into 2 × 2 pieces), if directly carrying out LBP processing since two-dimensional time-domain grayscale image is larger to picture, then cannot Local problem's feature of picture kind is extracted well.Whole picture figure can be carried out piecemeal by us, if the width of dimension of picture and height Degree is 2 multiple, then two dimensional gray figure can be blocked into 2 × 2 (as shown in Figure 3)；If the width of resolution ratio and height are 3 Multiple, then two dimensional gray figure can be blocked into 3 × 3 ... and so on.Piecemeal is more to make discrimination higher, but corresponding Operation time will will increase.
Above-mentioned gray level image is divided region and used based on unified local binary patterns by step 3Operator carries out special Sign, which is extracted, obtains each statistic histogram for dividing region, then the statistic histogram that each division region obtains is combined, and will Statistic histogram after combination is normalized, and makes its scope control in [0,1], obtains the statistic histogram of whole picture figure (such as Fig. 4).
The initial apparatus of LBP handles local image contrast, is easily understood due to possessing theory, and arithmetic speed is fast, and coding is easy The advantages of realization.Attached drawing 2 is referred to, enabling image kind pixel is pi, with surrounding 8 field (p1-p8) one template of composition. By 8 field pixel values respectively with center pixel piCompare, if difference is greater than 0, enabling difference state is 1, if enabling difference less than 0 State of value is 0.The decimal system is converted by binaryzation feature obtained according to certain method, i.e. the LBP of this center pixel is special Sign.The study found that only a small number of partial modes describe the important model of image texture, the probability occurred reaches 90% or more, This mode is exactly " More General Form ".If binary system is regarded as a circle, More General Form be exactly in string from 0 to 1 and from 1 to 0 Conversion be no more than mode twice, mark(0 transformation), 01110000 (turn for 2 times for example, mode 00000000 Become), 11001111 (2 times transformation) be More General Form, referred to as consistent LBP.And 11001001 (4 transformations), 01010011 (6 Secondary transformation) it is not More General Form, referred to as non-uniform LBP.With P=8, for R=1, LBPP,RMode have 256 kinds, andMode at most only have 59 kinds.ThusThe fault message in two-dimension vibration time domain grayscale image is not only saved, and Further dimensionality reduction.
The statistic histogram is input to DBN network by step 4, and preceding to self study by DBN network progress Processing and the backpropagation processing declined based on gradient, the model parameter of DBN network shown in adjustment obtain the DBN that training is completed Network, DBN network model are as shown in Figure 2.Depth confidence network is substantially by multiple limited Boltzmann machines (RBM) and one layer Multi-layered perception neural networks made of having the counterpropagation network (BPNN) of supervision to stack, each limited Boltzmann machine is by can It is formed depending on layer and hidden layer, includes 3 layers of RBM in the present embodiment, each layer of parameter is as follows.
The first layer: the first visual layers v1For original input data and the first hidden layer h1Form first RBM (RBM1)；
The second layer: the first hidden layer h1As the second visual layers v2, and with the second hidden layer h2Form second RBM (RBM2)；
The third time: the second hidden layer h2As third visual layers v3, and with third layer hidden layer h3Form third RBM(RBM3)
It is connected between layers by weight w, it is mutually indepedent inside each layer, it is each hidden when the state of given visible element Hiding between layer state of activation is conditional sampling, at this point, the activation probability of j-th of hidden unit are as follows:
WhereinFor sigmoid activation primitive, θ={ wij,ai,bjBe RBM parameter, wijIndicate visible Connection weight between unit i and hidden unit j, aiIndicate the biasing of visual element i, bjIndicate the biasing of Hidden unit j, hj It is j-th of hidden unit state, viIt is i-th of visible element state, there are two types of their states: 0 or 1.Due to the knot of RBM Structure is symmetrically, when the state of given hidden unit, it is seen that the activation condition of unit is also conditional sampling, i.e., i-th can Depending on the activation probability of unit are as follows:
RBM is the stochastic neural net that an activation primitive is sigmoid function, is trained by way of iteration To parameter θ={ wij,ai,bjAs a result, and being fitted with given training data.Optimized parameter θ*It can be by training set Very big log-likelihood function be
In order to quickly calculate the log-likelihood gradient of RBM, weight and biasing can be obtained using the algorithm to sdpecific dispersion Parameter updates, and indicates are as follows:
Δωij=ε (< vihj>data-<vihj>recon)
Δai=ε (< vi>data-<vi>recon)
Δbi=ε (< hj>data-<hj>recon)
Wherein: ε is the learning rate of pre-training,<>dataFor mathematic expectaion defined in training data,<>reconAttach most importance to Mathematic expectaion defined in the model of structure.
BP network is a kind of classifier of supervised learning, can be classified to the feature vector that RBM pre-training obtains, and And play the role of the entire DBN parameter of fine tuning.The training of BP network is divided into two stages of propagated forward and back-propagating.
The propagated forward stage: input feature value is successively traveled to output layer, the classification results predicted.
The reversed fine tuning stage: the classification results that prediction obtains are compared to obtain error with standard markup information, will accidentally Difference successively returns backward, to realize the fine tuning of DBN parameter.
If DBN is stacked by l RBM altogether, initial sample x, the last layer output vector is ul(x)
Wherein, blFor the biasing of l layers of RBM, wlFor the weight of l layers of RBM, i-th of sample is folded under RBM through preceding to l layer heap After study, belong to classification yi, yiThe probability of ∈ (1,2 ..., c) is
V is parameter coefficient, and choosing classification corresponding to maximum probability is that Soft-max model determines classification.
L layers of error function express formulas
ρ is weight attenuation rate, λl={ wl,bl,cl,Vl, 1 { yi=k } it is logic display function, work as yiWhen=k, being worth is 1. Work as yiWhen ≠ k, be worth be for 0, m hidden layer neuron number, c be classification number.To seek error minimum value, risen using gradient Method seeks local derviation to parameter
Finely tune parameter
A is learning rate, and so on, l layers of relevant parameter are arrived in fine tuning 1.
Step 5 utilizes the DBN net of obtained training completion when identifying practical multi-load rolling bearing fault mode Network does same treatment to the vibration signal and above-mentioned known state of the rolling bearing of unknown state, i.e., corresponding two-dimensional time-domain ash Spend figureStatistic histogram is handled, so that it is determined that the fault type of practical rolling bearing.
When it is implemented, technical solution of the present invention can realize automatic running process based on computer software technology, can also adopt Corresponding system is realized with modular mode.The embodiment of the present invention provides a kind of based on local binary patterns and depth confidence network Rolling bearing forecasting system, including following module:
Preprocessing module, the vibration signal for the one-dimensional rolling bearing to known state pre-process, and obtain known The two-dimensional time-domain grayscale image of the rolling bearing of state；
Image division module, for the two-dimensional time-domain gray level image of the rolling bearing of obtained known state to be carried out region It divides, division principle are as follows: take the common divisor of picture traverse and height respectively；
Characteristic extracting module is used for above-mentioned gray level image to be divided region based on unified local binary patternsOperator carries out feature extraction, obtains each statistic histogram for dividing region, then the statistics histogram that each division region is obtained Figure is combined, and obtains the statistic histogram of whole picture figure, and statistic histogram is normalized；
DBN network training module, for the statistic histogram of entire image to be input to DBN network (depth confidence net Network), and the backpropagation for handling before carrying out to DBN network to self study and being declined based on gradient is handled, and the mould of DBN network is adjusted Shape parameter obtains the DBN network of training completion；
Failure modes module is used for when identifying practical multi-load rolling bearing fault mode, first to unknown state The vibration signal of rolling bearing carries out the processing of step 1 to three, obtains its corresponding statistic histogram, then using acquired Training complete DBN network classify, so that it is determined that the fault type of practical rolling bearing.
Each module specific implementation can be found in corresponding steps, and the present invention not writes.
Embodiment
1, initial data prepares
The present embodiment is illustrated with Xi Chu university of U.S. bearing data instance based on local binary patterns and depth confidence network Rolling bearing fault diagnosis implementation method.
The rolling bearing test platform includes one 2 horsepowers of motor (left side) (1hp=746w), a torque sensor (centre), a power meter and control electronics.Use inner ring, outer ring and rolling of the spark erosion technique on rolling bearing Single Point of Faliure is respectively arranged on kinetoplast, fault diameter distinguishes 0.007,0.014,0.021 inch.The four SKF bearing of bearing used. The experimental bench includes drive end bearing and fan end bearing, and the sensor of acceleration is separately mounted to electric machine casing and driving end and wind Fan the position at 12 o'clock of end.Vibration signal is that the DAT logger in general 6 channel acquires, and the sample frequency of digital signal is every Second 12000 points, drive end bearing fault data sample rate are 48000 points each second, testing stand schematic diagram such as Fig. 6 institute Show.
The present embodiment to drive end (DE) bearing as research object, select fault diameter for 7mils, 14mils and 21mils.IR07, IR14, IR21 are indicated to the failure shape of rolling bearing inner ring lesion diameter 7mils, 14mils and 21mils Condition；And so on OR07, OR14 and OR21 respectively represent housing washer lesion diameter 7mils, 14mils and 21mils Fault state；B07, B14 and B21 respectively represent the failure of rolling bearing rolling element lesion diameter 7mils, 14mils and 21mils Situation；N represents normal condition.Every kind of state has tri- kinds of 0hp, 1hp and 2hp loads.With 3 kinds load under, construct 3 kinds of data set A, The each data set of B and C includes 10 kinds of fault categories, each available 50 samples of fault category, and each sample length is 2048.In view of actual conditions, same fault category may correspond to a variety of loads, and comprehensive A-C obtains data set D, random to select The 60% of each fault category sample is taken to be used as training, it is remaining as test.Specific bearing data can refer to table 1:
1 bearing data of table
The major parameter of model is as shown in table 2, and DBN model parameter is as shown in table 3, different block count test results Such as table 4；
Table 2Model parameter
Gray scale dimension of picture Piecemeal (i=1,2,4,5) P adjacent pixel points The R radius of neighbourhood
560×420 i×i 8 1
3 DBN model parameter of table
Hidden layer number of nodes Learning rate Momentum The number of iterations
100-100 0.0001 0.9 3000
The different block counts of table 4 identify situation
As shown in Table 4, with the increase of block count, identify that accurate rate increases accordingly, but runing time is consequently increased.For Validity of the verifying method of the invention in precision aspect, is compared, i.e., using two kinds of widely used sorting algorithms Counterpropagation network (BPNN) and support vector machines (SVM).In addition, we use mutually isostructural in order to protrude the validity of DBN BPNN or SVM replaces DBN to carry out the identification of similar structures.BPNN includes 3 hidden layers, and each layer includes 10 neurons. The kernel function of support vector machines uses radial basis function, and kernel function and penalty factor are automatically selected by cross validation method.Consider Diagnostic result in table 4 the time required to block count, discrimination and operation, choosing block count is 4 × 4, remaining parameter and above-mentioned phase Together, test result is as shown in table 5:
Table 5Model compares diagnostic result list
The experimental result of the method for the present invention case is analyzed, available to draw a conclusion:
Feature extraction by LBP to two-dimension vibration time domain grayscale image can obtain having for two-dimension vibration time domain grayscale image Textural characteristics, and dimensionality reduction is carried out to image.By testing it can be found that the Local textural feature extracted makes DBN can Still to keep higher diagnostic accuracy under multi-load.
The present invention provides a kind of rolling bearing prediction side based on More General Form local binary patterns and depth confidence network One-dimensional bearing vibration time-domain signal is changed into two-dimensional time-domain grayscale image first by method, can be suitably by picture piecemeal, after piecemeal The method for being utilized respectively unified local binary patterns carries out Local textural feature extraction to two-dimensional time-domain grayscale image, and by image Carry out dimensionality reduction.Obtained statistic histogram is merged, later as the input of DBN, by the further self study of DBN, Feature extraction simultaneously carries out fault category classification, compares by experimental result, demonstrates the method and not only loads the lower axis of rolling to single The malfunction held has accurate identification, and the case where be equally applicable to multi-load, more meets actual working condition.
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.

Claims (8)

1. a kind of rolling bearing prediction technique based on local binary patterns and depth confidence network, which is characterized in that including such as Lower step:
Step 1 pre-processes the vibration signal of the one-dimensional rolling bearing of known state, obtains the axis of rolling of known state The two-dimensional time-domain grayscale image held；
The two-dimensional time-domain gray level image of the rolling bearing of obtained known state is carried out region division, division principle by step 2 Are as follows: the common divisor of picture traverse and height is taken respectively；
Above-mentioned gray level image is divided region and used based on unified local binary patterns by step 3Operator carries out feature and mentions It takes, obtains each statistic histogram for dividing region, then the statistic histogram that each division region obtains is combined, obtain whole picture The statistic histogram of figure, and statistic histogram is normalized；
The statistic histogram of entire image is input to DBN network (depth confidence network) by step 4, and is carried out to DBN network Forward direction self study processing and the backpropagation processing declined based on gradient, the model parameter of adjustment DBN network obtain training completion DBN network；
Step 5, when identifying practical multi-load rolling bearing fault mode, the vibration to the rolling bearing of unknown state first Signal carries out the processing of step 1 to three, obtains its corresponding statistic histogram, the DBN then completed using obtained training Network is classified, so that it is determined that the fault type of practical rolling bearing.
2. a kind of rolling bearing prediction side based on local binary patterns and depth confidence network according to claim 1 Method, it is characterised in that: in the step 1, place is normalized in the vibration signal of the one-dimensional rolling bearing to known state Reason makes its scope control in [0,1], the vibration signal after normalization is converted to two dimensional gray time-domain diagram later, picture Amplitude is set as [0-1]；It is required that the vibration signal length for being converted into the one-dimensional rolling bearing of two dimensional gray time-domain diagram should meet greatly It turns around the number of the acquired data point of sensor in shaft；The resolution ratio of two dimensional gray figure scale having the same.
3. a kind of rolling bearing prediction side based on local binary patterns and depth confidence network according to claim 1 Method, it is characterised in that: in the step 2, region division is carried out to the two-dimensional time-domain gray level image of the rolling bearing of known state Specific Principles are as follows: divided according to the resolution sizes of two dimensional gray figure, if the width of resolution ratio and height be n times Number, then be blocked into n × n for two dimensional gray figure；And the more meeting discriminations of piecemeal are higher, but operation time will will increase.
4. a kind of rolling bearing prediction side based on local binary patterns and depth confidence network according to claim 1 Method, it is characterised in that: DBN network described in step 4 is substantially to have prison by multiple limited Boltzmann machines (RBM) and one layer Multi-layered perception neural networks made of the counterpropagation network (BPNN) superintended and directed stacks, specific structure is as follows,
Each limited Boltzmann machine is made of visual layers and hidden layer, is connected between layers by weight w, inside each layer It independently of each other, is conditional sampling between each hidden layer state of activation, at this point, j-th hidden when the state of given visible element Hide the activation probability of unit are as follows:
WhereinFor sigmoid activation primitive, θ={ wij, ai, bjBe RBM parameter, wijIndicate visible element Connection weight between i and hidden unit j, aiIndicate the biasing of visual element i, bjIndicate the biasing of Hidden unit j, hjIt is jth A hidden unit state, viIt is i-th of visible element state, there are two types of their states: 0 or 1；Since the structure of RBM is Symmetrically, when the state of given hidden unit, it is seen that the activation condition of unit is also conditional sampling, i.e., i-th visual single The activation probability of member are as follows:
RBM is the stochastic neural net that an activation primitive is sigmoid function, is trained and is joined by way of iteration Number θ={ wij, ai, bjAs a result, and be fitted with given training data, optimized parameter θ * can pass through the pole on training set Big log-likelihood function is
In order to quickly calculate the log-likelihood gradient of RBM, the parameter of weight and biasing can be obtained using the algorithm to sdpecific dispersion It updates, indicates are as follows:
Δωij=ε (< vihj>data-<vihj>recon)
Δai=ε (< vi>data-<vi>recon)
Δbi=ε (< hj>data-<hj>recon)
Wherein: ε is the learning rate of pre-training,<>dataFor mathematic expectaion defined in training data,<>reconFor reconstruct Mathematic expectaion defined in model；
BP network is a kind of classifier of supervised learning, can be classified to the feature vector that RBM pre-training obtains, and is risen To the effect for finely tuning entire DBN parameter, the training of BP network is divided into two stages of propagated forward and back-propagating；
The propagated forward stage: input feature value is successively traveled to output layer, the classification results predicted；
The reversed fine tuning stage: classification results that prediction obtains are compared to obtain error with standard markup information, by error by Layer returns backward, to realize the fine tuning of DBN parameter；
If DBN is stacked by l RBM altogether, initial sample x, the last layer output vector is ul(x),
Wherein, blFor the biasing of l layers of RBM, wlFor the weight of l layers of RBM, i-th of sample learns under RBM through preceding fold to l layer heap Afterwards, belong to classification yi, yi∈ (1,2 ..., probability c) be
V is parameter coefficient, and choosing classification corresponding to maximum probability is that Soft-max model determines classification；
L layers of error function express formulas
ρ is weight attenuation rate, λl={ wl, bl, cl, Vl, 1 { yi=k } it is logic display function, work as yiWhen=k, being worth is 1, works as yi When ≠ k, be worth be for 0, m hidden layer neuron number, c be classification number；To seek error minimum value, using gradient rise method, Local derviation is asked to parameter
Finely tune parameter
A is learning rate, and so on, l layers of relevant parameter are arrived in fine tuning 1.
5. a kind of rolling bearing forecasting system based on local binary patterns and depth confidence network, which is characterized in that including such as Lower module:
Preprocessing module, the vibration signal for the one-dimensional rolling bearing to known state pre-process, and obtain known state Rolling bearing two-dimensional time-domain grayscale image；
Image division module is drawn for the two-dimensional time-domain gray level image of the rolling bearing of obtained known state to be carried out region Point, division principle are as follows: take the common divisor of picture traverse and height respectively；
Characteristic extracting module is used for above-mentioned gray level image to be divided region based on unified local binary patterns Operator carries out feature extraction, obtains each statistic histogram for dividing region, then by each division obtained statistic histogram in region into Row combination, obtains the statistic histogram of whole picture figure, and statistic histogram is normalized；
DBN network training module, for the statistic histogram of entire image to be input to DBN network (depth confidence network), and The backpropagation for handling before carrying out to DBN network to self study and being declined based on gradient is handled, the model ginseng of adjustment DBN network Number obtains the DBN network of training completion；
Failure modes module is used for when identifying practical multi-load rolling bearing fault mode, first to the rolling of unknown state The vibration signal of bearing carries out the processing of step 1 to three, obtains its corresponding statistic histogram, then utilizes obtained instruction Practice the DBN network completed to classify, so that it is determined that the fault type of practical rolling bearing.
6. according to claim 5 a kind of based on the rolling bearing of local binary patterns and depth confidence network prediction system System, it is characterised in that: in the preprocessing module, the vibration signal of the one-dimensional rolling bearing to known state is normalized Processing makes its scope control in [0,1], the vibration signal after normalization is converted to two dimensional gray time-domain diagram, picture later Amplitude be set as [0-1]；It is required that being converted into the vibration signal length of the one-dimensional rolling bearing of two dimensional gray time-domain diagram should meet It turns around the number of the acquired data point of sensor greater than shaft；The resolution ratio of two dimensional gray figure scale having the same.
7. according to claim 5 a kind of based on the rolling bearing of local binary patterns and depth confidence network prediction system System, it is characterised in that: in described image division module, area is carried out to the two-dimensional time-domain gray level image of the rolling bearing of known state The Specific Principles that domain divides are as follows: divided according to the resolution sizes of two dimensional gray figure, if the width of resolution ratio and height are equal For the multiple of n, then two dimensional gray figure is blocked into n × n；And the more meeting discriminations of piecemeal are higher, but operation time will will increase.
8. a kind of rolling bearing prediction side based on local binary patterns and depth confidence network according to claim 5 Method, it is characterised in that: DBN network described in DBN network training module be substantially by multiple limited Boltzmann machines (RBM) and One layer has multi-layered perception neural networks made of counterpropagation network (BPNN) stacking of supervision, and specific structure is as follows,
Each limited Boltzmann machine is made of visual layers and hidden layer, is connected between layers by weight w, inside each layer It independently of each other, is conditional sampling between each hidden layer state of activation, at this point, j-th hidden when the state of given visible element Hide the activation probability of unit are as follows:
WhereinFor sigmoid activation primitive, θ={ wij, ai, bjBe RBM parameter, wijIndicate visible element Connection weight between i and hidden unit j, aiIndicate the biasing of visual element i, bjIndicate the biasing of Hidden unit j, hjIt is jth A hidden unit state, viIt is i-th of visible element state, there are two types of their states: 0 or 1；Since the structure of RBM is Symmetrically, when the state of given hidden unit, it is seen that the activation condition of unit is also conditional sampling, i.e., i-th visual single The activation probability of member are as follows:
RBM is the stochastic neural net that an activation primitive is sigmoid function, is trained and is joined by way of iteration Number θ={ wij, ai, bjAs a result, and be fitted with given training data, optimized parameter θ * can pass through the pole on training set Big log-likelihood function is
In order to quickly calculate the log-likelihood gradient of RBM, the parameter of weight and biasing can be obtained using the algorithm to sdpecific dispersion It updates, indicates are as follows:
Δωij=ε (< vihj>data-(vihj>recon)
Δai=ε ((vi>data-<vi>recon)
Δbi=ε (< hj>data-<hj>recon)
Wherein: ε is the learning rate of pre-training,<>dataFor mathematic expectaion defined in training data,<>reconFor reconstruct Mathematic expectaion defined in model；
BP network is a kind of classifier of supervised learning, can be classified to the feature vector that RBM pre-training obtains, and is risen To the effect for finely tuning entire DBN parameter, the training of BP network is divided into two stages of propagated forward and back-propagating；
The propagated forward stage: input feature value is successively traveled to output layer, the classification results predicted；
The reversed fine tuning stage: classification results that prediction obtains are compared to obtain error with standard markup information, by error by Layer returns backward, to realize the fine tuning of DBN parameter；
If DBN is stacked by l RBM altogether, initial sample x, the last layer output vector is ul(x),
Wherein, blFor the biasing of l layers of RBM, wlFor the weight of l layers of RBM, i-th of sample learns under RBM through preceding fold to l layer heap Afterwards, belong to classification yi, yi∈ (1,2 ..., probability c) be
V is parameter coefficient, and choosing classification corresponding to maximum probability is that Soft-max model determines classification；
L layers of error function express formulas
ρ is weight attenuation rate, λl={ wl, bl, cl, Vl, 1 { yi=k } it is logic display function, work as yiWhen=k, being worth is 1, works as yi When ≠ k, be worth be for 0, m hidden layer neuron number, c be classification number；To seek error minimum value, using gradient rise method, Local derviation is asked to parameter
Finely tune parameter
A is learning rate, and so on, l layers of relevant parameter are arrived in fine tuning 1.
CN201810742624.9A 2018-07-09 2018-07-09 A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network Pending CN108932499A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810742624.9A CN108932499A (en) 2018-07-09 2018-07-09 A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810742624.9A CN108932499A (en) 2018-07-09 2018-07-09 A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network

Publications (1)

Publication Number Publication Date
CN108932499A true CN108932499A (en) 2018-12-04

Family

ID=64447866

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810742624.9A Pending CN108932499A (en) 2018-07-09 2018-07-09 A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network

Country Status (1)

Country Link
CN (1) CN108932499A (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109519340A (en) * 2018-12-27 2019-03-26 中国船舶重工集团海装风电股份有限公司 A kind of wind turbine generator drive system method for diagnosing faults
CN109635931A (en) * 2018-12-14 2019-04-16 吉林大学 A kind of equipment running status evaluation method based on depth conviction net
CN109753882A (en) * 2018-12-13 2019-05-14 金陵科技学院 Sign Language Recognition Method based on depth confidence network and multi-mode feature
CN110246134A (en) * 2019-06-24 2019-09-17 株洲时代电子技术有限公司 A kind of rail defects and failures sorter
CN111080607A (en) * 2019-12-12 2020-04-28 哈尔滨市科佳通用机电股份有限公司 Rolling bearing oil slinging fault detection method based on image recognition
CN111209891A (en) * 2020-01-15 2020-05-29 佛山科学技术学院 Deep neural network-based bearing working condition detection method and system
CN111553112A (en) * 2020-03-16 2020-08-18 广西电网有限责任公司电力科学研究院 Power system fault identification method and device based on deep belief network
CN112150497A (en) * 2020-10-14 2020-12-29 浙江大学 Local activation method and system based on binary neural network
CN112282739A (en) * 2020-11-18 2021-01-29 中国石油天然气集团有限公司 Method for identifying debris scatterers in borehole diameter measurement while drilling
CN112507479A (en) * 2020-11-19 2021-03-16 电子科技大学 Oil drilling machine health state assessment method based on manifold learning and softmax

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106769048A (en) * 2017-01-17 2017-05-31 苏州大学 Adaptive deep confidence network bearing fault diagnosis method based on Nesterov momentum method
CN107578007A (en) * 2017-09-01 2018-01-12 杭州电子科技大学 A kind of deep learning face identification method based on multi-feature fusion
CN107729890A (en) * 2017-11-30 2018-02-23 华北理工大学 Face identification method based on LBP and deep learning

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106769048A (en) * 2017-01-17 2017-05-31 苏州大学 Adaptive deep confidence network bearing fault diagnosis method based on Nesterov momentum method
CN107578007A (en) * 2017-09-01 2018-01-12 杭州电子科技大学 A kind of deep learning face identification method based on multi-feature fusion
CN107729890A (en) * 2017-11-30 2018-02-23 华北理工大学 Face identification method based on LBP and deep learning

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
单外平 等: ""基于深度信念网络的信号重构与轴承故障识别"", 《电子设计工程》 *
宋嫱嫱: ""基于稀疏表示的鲁棒性人脸识别研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *
崔明利: ""基于多尺度几何分析的旋转机械故障诊断研究"", 《中国优秀硕士学位论文全文数据库 工程科技II辑》 *
杨印生: "《经济系统定量分析方法》", 30 April 2001, 吉林科学技术出版社 *
陈海周: ""旋转机械传动件微弱故障融合诊断方法研究"", 《中国博士学位论文全文数据库 工程科技Ⅱ辑》 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109753882A (en) * 2018-12-13 2019-05-14 金陵科技学院 Sign Language Recognition Method based on depth confidence network and multi-mode feature
CN109635931A (en) * 2018-12-14 2019-04-16 吉林大学 A kind of equipment running status evaluation method based on depth conviction net
CN109519340A (en) * 2018-12-27 2019-03-26 中国船舶重工集团海装风电股份有限公司 A kind of wind turbine generator drive system method for diagnosing faults
CN110246134A (en) * 2019-06-24 2019-09-17 株洲时代电子技术有限公司 A kind of rail defects and failures sorter
CN111080607A (en) * 2019-12-12 2020-04-28 哈尔滨市科佳通用机电股份有限公司 Rolling bearing oil slinging fault detection method based on image recognition
CN111209891A (en) * 2020-01-15 2020-05-29 佛山科学技术学院 Deep neural network-based bearing working condition detection method and system
CN111553112A (en) * 2020-03-16 2020-08-18 广西电网有限责任公司电力科学研究院 Power system fault identification method and device based on deep belief network
CN112150497A (en) * 2020-10-14 2020-12-29 浙江大学 Local activation method and system based on binary neural network
CN112282739A (en) * 2020-11-18 2021-01-29 中国石油天然气集团有限公司 Method for identifying debris scatterers in borehole diameter measurement while drilling
CN112282739B (en) * 2020-11-18 2024-04-30 中国石油天然气集团有限公司 Method for identifying rock debris scatterer in measurement while drilling diameter
CN112507479A (en) * 2020-11-19 2021-03-16 电子科技大学 Oil drilling machine health state assessment method based on manifold learning and softmax
CN112507479B (en) * 2020-11-19 2022-03-15 电子科技大学 Oil drilling machine health state assessment method based on manifold learning and softmax

Similar Documents

Publication Publication Date Title
CN108932499A (en) A kind of rolling bearing prediction technique and system based on local binary patterns and depth confidence network
Lu et al. Deep model based domain adaptation for fault diagnosis
WO2021134871A1 (en) Forensics method for synthesized face image based on local binary pattern and deep learning
CN112763214B (en) Rolling bearing fault diagnosis method based on multi-label zero-sample learning
CN106980822B (en) A kind of rotary machinery fault diagnosis method based on selective ensemble study
CN105184312B (en) A kind of character detecting method and device based on deep learning
Zhang et al. Infrared image segmentation for photovoltaic panels based on Res-UNet
Ma et al. Cross-domain meta learning fault diagnosis based on multi-scale dilated convolution and adaptive relation module
Gao et al. ASM1D-GAN: An intelligent fault diagnosis method based on assembled 1D convolutional neural network and generative adversarial networks
CN109063719B (en) Image classification method combining structure similarity and class information
CN104992223A (en) Intensive population estimation method based on deep learning
Arkin et al. A survey of object detection based on CNN and transformer
CN113988126A (en) Rolling bearing fault diagnosis method based on less-label data feature migration
CN108549912A (en) A kind of medical image pulmonary nodule detection method based on machine learning
CN106127198A (en) A kind of image character recognition method based on Multi-classifers integrated
CN106022254A (en) Image recognition technology
Ren et al. Learning convolutional nonlinear features for k nearest neighbor image classification
CN110111365B (en) Training method and device based on deep learning and target tracking method and device
Liu et al. A image segmentation algorithm based on differential evolution particle swarm optimization fuzzy c-means clustering
Darwish Soft computing applied to the build of textile defects inspection system
Zhang et al. Intelligent fault diagnosis of rolling bearing using the ensemble self‐taught learning convolutional auto‐encoders
CN105809200A (en) Biologically-inspired image meaning information autonomous extraction method and device
CN108898157B (en) Classification method for radar chart representation of numerical data based on convolutional neural network
Yap et al. Compressing and improving fuzzy rules using genetic algorithm and its application to fault detection
CN116975622A (en) Training method and device of target detection model, and target detection method and device

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20181204